Test that the service runs and uses our configured (TCP) port.
This should prevent https://github.com/NixOS/nixpkgs/issues/475043
("nixos/scanservjs: settings (host/port) ignored in v3.0.4") from
happening again.
(cherry picked from commit 32157e1c52)
scanservjs has been ignoring the configuration options set by the NixOS
module since the v3.0.4 upgrade. This fixes it.
Create dirs/files in the NixOS module as needed by the new scanservjs
build.
Fixes: 7432e50d62 ("scanservjs: 2.27.1 -> 3.0.4")
Fixes: https://github.com/NixOS/nixpkgs/issues/475043 ("nixos/scanservjs: settings (host/port) ignored in v3.0.4")
Co-authored-by: Bjørn Forsman <bjorn.forsman@gmail.com>
(cherry picked from commit 008e87e5c2)
rebar3 itself actually compiles, it's just the tests that are failing.
Patch to disable another warning until upstream fixes the tests.
(cherry picked from commit 09d3443b49)
MLX backends require xcrun metal/metalib which are unavailable in the
Nix sandbox. Pass -DOLLAMA_MLX_BACKENDS="" to cmake to skip the
Metal toolchain check. Core Metal/llama.cpp backend is unaffected.
(cherry picked from commit b195b40fb0)
Addresses two review comments — different reviewers, same direction:
* @SuperSandro2000 on the original PR #528150 (line 118 of
pkgs/by-name/ol/ollama/package.nix): "Why are we not using tag if
there is a tag?". Drop the SHA + `# tag b9509` shape that required
reviewers to cross-check the SHA matched the inline comment.
* @prusnak on this follow-up #528374: "I think we should use `tag`
instead of `rev`." `fetchFromGitHub`'s dedicated `tag` field is
preferred over `rev = "<tag-string>"` — same git object, same
tarball, same hash, but signals intent at the schema level and
renders into the store path nicely.
* Line 113 prose comment (the other half of SuperSandro's review on
#528150): "We should not repeat the version in this comment and
have to update that, too, on every update." Drop `currently b9509`
from the prose; the remaining text explains the mechanism (tracks
upstream's `LLAMA_CPP_VERSION` file, pre-staged because
FetchContent can't network in the sandbox) without naming a
specific version that would drift on every bump.
Pure source-readability change. `nix-build -A ollama --no-out-link`
on x86_64-linux produces the same store path as master, confirming
the tag resolves to the same git object → same tarball → same hash
→ no rebuild impact.
The same lines exist on `release-26.05` via the original backport
#528272; happy to mirror this fix there if you apply the
`backport release-26.05` label.
Assisted-by: Claude <noreply@anthropic.com>
(cherry picked from commit 1e9315fd0c)
